Feature importance guided autoencoder for dimensionality reduction in intrusion detection systems.

Area of Science:

• Cybersecurity
• Machine Learning
• Network Security

Background:

• Intrusion detection systems (IDS) are crucial for network protection.
• Dimensionality reduction improves machine learning-based IDS effectiveness.
• Existing methods may not fully optimize feature selection for IDS.

Purpose of the Study:

• To propose an effective dimensionality reduction technique for IDS.
• To enhance the accuracy and performance of machine learning-based IDS.
• To introduce a novel feature importance method for network intrusion data.

Main Methods:

• Developed a feature importance-based autoencoder (FI-AE).
• Introduced a novel one-versus-all (OVA) feature importance method using random forests.
• Trained an autoencoder with a weighted loss function based on OVA feature importance.
• Applied the autoencoder for dimensionality reduction, followed by a random forest classifier.

Main Results:

• The FI-AE model effectively reduced feature dimensionality in benchmark datasets.
• The random forest classifier achieved higher accuracy and F1-score with the FI-AE reduced data.
• Outperformed previous dimensionality reduction techniques on NSL-KDD, UNSW-NB15, and CIC-IDS2017 datasets.

Conclusions:

• The proposed FI-AE technique is an effective dimensionality reduction method for IDS.
• Integrating OVA feature importance with autoencoders enhances IDS performance.
• This approach offers a promising direction for improving network security through advanced machine learning.